Traffic-Tran: A Parallel Multi-encoder Structure for Cellular Traffic Prediction

Wireless cellular traffic prediction is a critical research topic for the realization of intelligent communications. The high nonlinearities and mutability of wireless cellular network traffic bring great challenges to accurate prediction. Due to the lack of dynamic spatio-temporal correlation modeling ability and complex network structure, the existing prediction methods cannot meet the requirements of accuracy and complexity in real scenes. By generating time series data for network traffic of a single grid, and spatial series data for network traffic of all grids with the same timestamp, this paper proposes a multi-encoder structure named “Traffic-Tran”, which learns sequence correlation independently and in parallel by multiple network units. Meanwhile, in order to improve the recognition ability of multi-encoder feature information, an information supplement method is proposed. In addition, the design of sampling output module realizes the parallel multi-step flow prediction, which enlarges the application range of the model. Experimental results on a large real dataset verify the effectiveness of Traffic-Tran. The model complexity of Traffic-Tran is greatly reduced, with less memory usage and shorter runtime than other models. Under the premise of the same predictive performance, the number of training parameters of Traffic-Tran is reduced by 44.9%.